A machine-learning surrogate model for optimising photovoltaic-thermal deployment in complex urban morphologies

Peer-reviewed paper 8-244-26

Photovoltaic-thermal (PVT) systems can simultaneously generate electricity and recover usable heat, significantly improving the utilisation of limited urban roof space. However, their performance in dense cities is strongly influenced by urban morphology and installation geometry, factors not adequately captured by conventional photovoltaic forecasting approaches. This study proposes a simulation-trained machine learning (ML) framework designed to optimise PVT layouts by evaluating key parameters, including panel density (inter-panel spacing), angles of inclination, and orientation, to achieve optimal electrical and thermal generation across varying urban environments.

A parametric dataset of 1,575 installation scenarios was generated using GIS-based urban modelling and building performance simulations across three morphologically distinct districts of Tehran. Artificial Neural Network (ANN) and Random Forest (RF) models were trained to predict electricity generation, hot water production, and associated carbon reduction (CR). Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE) and mean absolute error (MAE) metrics.

The RF model achieved high predictive accuracy (R² = 0.91–0.99) with shorter training times than the ANN. Sensitivity analysis revealed that urban form and installation geometry (panel spacing and tilt) exert a stronger influence on combined energy yield than several climatic predictors. By replacing computationally intensive simulations with a data-driven surrogate, the proposed framework enables rapid design-space exploration and provides a robust tool for informed PVT deployment in energy-constrained urban environments.